When chatbots first debuted decades ago, companies tried their hand at creating bots — especially for customer service. Thus began a giant experiment in user experience. Chatbots haven’t always fared well on that front, but recent research suggests they are not annoying consumers as much as they used to.

Turns out, chatbot design — how you go about building your bot — matters. But what actually goes into making a chatbot successful? 

To learn more about the principles of chatbot design thinking, we sat down with Conversational AI expert Casey Phillips who helped us understand how to approach creating bots and navigate the challenges that may arise in the process.

Chatbot design: Where do you begin?

First, we need to clarify what chatbot design means. It’s more than just training chatbots to answer predetermined questions – it’s about making them conversational, giving them a personality — a voice — and the “brains” to actually converse with humans. Phillips explained:

“The art of successful conversational design is the practice of creating a conversational experience that provides value and benefits to users that they won't get from a traditional, non-conversational experience. For example, a well-designed chatbot could help users get answers to FAQs faster than they could from a traditional search-driven experience.”

On the other hand, he added, if a chatbot causes it to take longer for users to get answers to FAQs then its design would be inadequate. “As a chatbot designer, you need to either question and rethink your design or maybe even question whether or not you have a worthy use case for a chatbot altogether,” he said. “The key is to have a goal or outcome you're looking to accomplish and to let that drive how you execute the design of your chatbot.”

While use cases may differ, the steps to design a chatbot (and write chatbot scripts) are similar for every bot:

1. Define what your chatbot will do (e.g. who will it serve, customers or prospects?)
2. Think about your chatbot’s personality (e.g. serious or cheeky? Peppy or matter-of-fact? A robot or a human avatar?)
3. List the topics your chatbot will cover (e.g. support, pricing, delivery updates)
4. Brainstorm user questions/statements for each topic
5. Train your chatbot to understand variations of questions/statements (e.g. “I want to talk to Sales”, “Can I talk to Sales?”)
6. Write your chatbot’s answers or prompts
7. Build out sequences to guide people through their conversation with your bot
8. Analyze the chatbot’s performance and train it some more

While this may seem like a straightforward process, it actually involves a lot of planning. But, don’t worry: even if you’re not an expert, you can still build a functional chatbot. And, it’s important to consider a few factors that can affect your chatbot’s performance.

First, check your expectations

One piece of advice Phillips offered is to accept that your bot can’t - and won’t - do it all. Even AIs like Siri, Cortana, and Alexa can’t do everything – and they’re much more advanced than your typical customer service bot. 

Also, attempting to answer every question under the sun poses a data challenge, Phillips said. Using data to train chatbots is key to building them right, so if your chatbot tries to answer everything, you’ll end up with too much data that will be difficult to keep up with. “Casting too wide a net will affect the performance of [your chatbot] because you'll just be overloading it with data and it’ll start to get confused,” said Phillips. “That’s where we see a lot of performance issues.”

So, check your expectations for chatbot design and make sure your team (and your customers) understand the capabilities of your chatbot.

Define your chatbot’s role

Pick a realistic business goal you want to achieve and target your bot toward that outcome. For example, the majority of chatbots offer support and troubleshoot frequently asked questions, because companies are often looking to automate repetitive processes and reduce their support costs. But this doesn’t mean your company needs a traditional support bot. You may, instead, need help with sales.

“If [for example], you’re a small real estate company, you probably don't have a lot of support questions,” said Phillips. “So there might not be a lot of value in creating a support bot. There might be more value to a chatbot for lead generation [since] the more leads you get from potential buyers or sellers, the more revenue you’ll drive.”

The more targeted a chatbot is, the better it works

You want people to get the information they need in an engaging way, and that’s only possible if you target your chatbot’s functions to what your audience needs— without diverting their attention to other topics or complicating the bot’s responses.

“The chatbots I’ve seen perform well are usually focused on one area of knowledge or questions – for example, filing taxes,” Phillips said. “If a chatbot helps users file their taxes, it should focus on answering the possible questions about that topic to offer the best possible experience.”

Often, you can add multiple functions to your chatbot. These might include clickable bubbles like ‘Support’, ‘Sales’, or ‘More information’ that guide visitors down a structured sequence. (More on that a little later).

So, to define your chatbot’s role, ask yourself: 

• What are my company’s business needs?
• How can I format my chatbot to converse with different types of users?
• What are the top questions my customers have that can be automated?

The operative phrase is “that can be automated.” It’s not wise to automate technical questions that require a lot of information from people, or are too tough to grasp. This will only frustrate your visitors. Think about areas where you can really improve customer experience by automating answers, and make sure you are offering something of value.

Give your chatbot a personality

Once you’re clear about what a chatbot should do, you can start thinking about “who” that chatbot will be. Will it be a humanoid with a real name and an avatar (kind of like Nadia, a bot developed for the Australian government)? Or will it be a smiling robot with antennas and a practical name like “SupportBot”? This is the first step in determining the personality of your bot.

Then, think about the language your bot should use. Usually, bots that use the idiosyncrasies of human conversation (like “Hm”, “What’s up?” or “LOL”) may be more engaging. But that should also depend on your chatbot use case – if you want a chatbot that will answer questions about taxation, you’ll probably give it a more serious tone of voice (and you’ll most likely avoid “LOL”).

Keep personality in mind when you build out your bot’s answers. It’s also good to consider human sentiment in each interaction, as Phillips himself has written about. For example, when the chatbot is helping a user with a minor or positive topic, like placing an order, it can speak in an upbeat tone and maybe even use humor. If, however, the bot is speaking to someone about a serious matter (e.g. filling an insurance claim), it’s better to keep its answers serious, too.

Consider the way people message

Making chatbots sound conversational is no easy feat. What people usually expect is that they’ll ask something and chatbots will understand it and reply in an appropriate way with the right tone. However, the intricacies of human communication might make interactions tricky for chatbots.

Here are two examples of challenges your chatbot might face:

1. Clarifying user statements

You can train chatbots to answer specific questions about a topic. You’ll usually collect feedback from your team and customers on the most common topics people ask about and try to come up with question variations and answers.

However, what happens when users don’t actually ask these questions verbatim, but instead say something generic? 

Here’s an example that Phillips brought up: imagine you have trained your chatbot to understand the sentences “I want to see the balances of my 401k” or “Can I enroll in 401k?” or “What are the terms and conditions of my 401k?” And then, someone comes along and asks, “I want help with my 401k.”

Unlike a human, the chatbot will probably find this statement too generic to understand. It needs to have a built-in way to disambiguate user statements – to find the meaning behind what is said. There are two ways you can resolve this when designing a chatbot:

• Populate a menu of 401k questions any time a person asks a question with the word “401k”. This is the easiest way since the message will be the same each time. But, this isn’t the greatest experience, because if someone does actually say, “I want to enroll in a 401k,” it’ll most likely be frustrating to scroll through a menu.

• Incorporate models and logic that will look at questions and match them to your database of existing answers. With this method, if the user asks something and there's only one answer that's relevant, the bot will always respond with that answer. If they ask something where several answers may be possible, the bot can either come back to the person and say “Hey, could you please elaborate with more details?” or respond with “Here's everything I can answer about 401k.”

2. Handling fragmented messages

Take a look at your most recent text messages with a friend or colleague. Chances are you’ll find that you often don’t send one long message to make your point, but multiple short ones that complete your thought when put together. For instance, see how a sentence is pieced together by the four bubbles in the screenshot below.

Most chatbots wouldn’t know how to handle a string of messages like this. They might try to process and respond to the user after each statement, which could lead to a frustrating user experience. The bot may respond to the first statement, and ask for more information—while all the information could have actually been given already, just in bits and pieces. “It's a pretty big limitation right now on how these systems work,” said Phillips.

How can you resolve this challenge in chatbot design? One possible solution is to set a delay to your chatbot’s responses. “The chatbot could wait maybe two or three seconds after the user has finished responding and they haven't typed anything, and you can try to group whatever the user said together,” Phillips said.  

But, according to Phillips, this might end up making the performance worse, hence why it’s not very common. That’s not only because of the delay, but also because the chatbot may be confused if users ask more than one question at the same time. Maybe the chatbot has a match for one question but not for the other. 

So you might be more successful in trying to resolve this by informing the user about what the chatbot can help them with and let them click on an option.

Find the balance between user-typed statements and clickable bubbles

Usually, if you only use the clickable bubbles or the ready prompts we mentioned in the previous section (‘Support’, ‘Sales’, ‘Exploring’), you will spare your bot the embarrassment of not understanding user statements. But, you’re also going to limit its capabilities. 

Phillips mentions that the best chatbots do a great job providing positive experiences both when users type their response and also when they click on buttons to go through a sequence. That’s because these bots cater to a wider audience with varying communication styles.

“Based on user research, there are people who would prefer a guided approach where the chatbots give them the potential response options, and there are also a lot of users who enjoy typing their question freely. As a designer, you need to work to make your chatbot good at both.”

It's all about using the right tech to build chatbots and striking a balance between free-form conversations and structured ones. Don’t be afraid to start an interaction with clickable responses to guide visitors down the right conversation path. But, try to make it possible for the chatbot to understand and reply to a user-typed response when needed by training it with specific questions variations.

Could the chatbots train the users too?

We’re talking about chatbots being trained from data generated by user behavior. But what if the chatbot could also serve to train the user on how to interact with AI tech?

Undoubtedly, consumers are becoming more and more familiar with chatbots. The first chatbots decades ago weren’t so popular because people weren’t used to messaging over the internet (most people didn’t even have a connection). But, as messaging has become an indispensable part of our lives, talking to digital beings has gotten easier.

And this journey toward learning to communicate with conversational AI will continue. A practical way to make the journey smoother is for chatbots to explicitly tell its users how to make their statements understood. For example, if someone is struggling, the chatbot can say something akin to “Try to format your answer in this way” or “Here are examples of questions I can answer.”

“I think that helps users who aren't familiar with chatbots [learn how] to converse with them. It helps them start to actually see into the world of chatbots and understand that, ‘Oh, I should just ask really short, concise questions, but also limit each of my questions to a single thought.’” 

This is how people will start getting used to speaking with chatbots. And, this is also a great way to set expectations. “It helps people understand that this bot was designed to answer targeted questions and there's a proper way to actually ask them.”

It’s all for your customers’ sake

At this point, you’re probably thinking that proper chatbot design takes time. And you’d be right – that’s why the roles of dedicated conversational designers have started growing, after all.

But, in the end, you’ll find it’s all worth it. Chatbots provide a number of benefits for business, and arguably, the biggest one is better customer experiences. In a world where customers expect more from businesses than ever before when it comes to good service, being able to resolve issues quickly or provide information 24/7 is a staple of modern customer support.

But, keep in mind that these benefits only come when the chatbot is good. If it doesn’t work as it should, it can have the opposite effect and tank your customer experience. In fact, the existence of bots that aren’t well-designed is why some people still don’t like them – both statistics and anecdotal evidence suggest that bad chatbots can make customer service worse than no chatbots at all. 

And that’s your cue

Thankfully, perceptions have been shifting, and that’s because there are chatbots coming out that are proving valuable. People are starting to have positive experiences and that means that they’re increasingly embracing chatbot technology. 

So, when you build your next bot, design it well. In a world that has just begun to understand the value of a ‘good’ chatbot, yours can be a critical competitive advantage if you pay attention to user experience. Phillips hit the nail on the head:

“You have to find moments, even small ones, that really impress users. Even simply knowing their location, like the weather app saying ‘Hey John, I see it's rainy in Seattle, hope you stay dry.’ These things are easy on the backend but they impress people, and help them overcome negative notions about chatbots and use them more.”